A bpmn-based lithology identification active learning method and system

Abstract

The application discloses a lithology identification active learning method and system based on BPMN, and the system comprises a BPMN process front-end module, a data set labeling module, a lithology identification active learning task module, a dynamic form module, a camunda process engine module, a model deployment module and a storage module. The application automatically trains a target on a server side based on a BPMN2.0 specification to realize a lithology identification function. The lithology identification active learning method based on the BPMN is used to increase the calculation of the credibility, the labeling candidate set extraction module and the labeling module on the original machine learning steps, improve the parameter learning, the model training and the complexity through the active learning mode of the machine, improve the model effect, and improve the work efficiency.

Description

Technical Field

[0001] This invention relates to the field of lithology identification, specifically a lithology identification active learning method and system based on BPMN. Background Technology

[0002] Lithology identification is an important basic geological research component in the drilling process. As the foundation for reservoir characteristic research and geological modeling, it provides a reliable basis for the study of spatial distribution in mining areas and is of great significance for the establishment of stratigraphic frameworks, the optimization of drilling engineering parameters, and the comprehensive evaluation of reservoirs.

[0003] Active learning is divided into a learning engine and a selection engine: the learning engine maintains the baseline classifier and uses supervised learning to learn from the labeled examples provided by the system to improve the performance of the classifier; the selection engine is responsible for running the sample selection algorithm, selecting an unlabeled example for experts to label, and adding the labeled example to the set of labeled examples, effectively discovering high-information samples in the training dataset without reducing classification accuracy.

[0004] BPMN is a standardized software process modeling notation. Using BPMN, the control flow and data flow of a software process can be clearly described, facilitating business process analysis and improvement, and enhancing work efficiency and understanding. Summary of the Invention

[0005] The purpose of this invention is to provide a lithology identification active learning method and system based on BPMN, which improves the accuracy and efficiency of lithology identification.

[0006] To achieve the above functions, this invention designs a lithology identification active learning method based on BPMN. For each target well log, the following steps S1-S5 are executed to construct a lithology identification active learning model, and then the lithology identification of each target well log is completed by applying the lithology identification active learning model:

[0007] Step S1: Based on the preset logging method, collect logging data from each target well, normalize the logging data of each target well to obtain each logging data sample, and construct a feature dataset x consisting of each logging data sample; combine the preset lithology labels corresponding to the preset lithology categories to construct a lithology label set y consisting of each lithology label.

[0008] Step S2: Randomly select n well logging data samples from the feature dataset x, and label each well logging data sample one by one according to each lithology label set y, and construct a sample dataset D from the labeled well logging data samples;

[0009] Step S3: Based on the CART decision tree algorithm, divide the decision tree according to lithology classification and construct the initial active learning model for lithology identification; take the sample dataset D constructed in step S2 as input and the lithology label corresponding to each well logging data sample as output, and train the initial active learning model for lithology identification a preset number of times.

[0010] Step S4: Randomly select n well logging data samples from the feature dataset x, input each well logging data sample into the lithology identification active learning initial model obtained in step S3, use the lithology label corresponding to each well logging data sample output by the lithology identification active learning initial model as pseudo label, assign each pseudo label to each well logging data sample one by one, and construct an unlabeled dataset U.

[0011] For the unlabeled dataset U, based on the query function and the preset confidence threshold, a candidate sample dataset L is constructed using well logging data samples with pseudo-label confidence lower than the preset confidence threshold. Based on the preset mapping rules, each well logging data sample in the candidate sample dataset L is labeled with a lithological label, and a labeled dataset E is constructed.

[0012] A pseudo-label dataset P is constructed using well logging data samples with pseudo-label credibility not lower than a preset credibility threshold;

[0013] Step S5: Input the sample dataset D obtained in step S2, the labeled dataset E obtained in step S4, and the pseudo-label dataset P into the initial active learning model for lithology identification, update it to obtain the active learning model for lithology identification, and apply the active learning model for lithology identification to complete the identification of lithology for each target well.

[0014] As a preferred technical solution of the present invention: the preset logging method in step S1 includes one or more of gamma-ray logging, sonic logging, density logging, compensated neutron logging, deep lateral logging, shallow lateral logging, and caliper logging; the lithology label includes acid rocks, coal, sandstone conglomerate, coarse sandstone, medium sandstone, fine sandstone, siltstone, and mudstone.

[0015] As a preferred technical solution of the present invention, the specific steps of step S3 are as follows:

[0016] Step S31: The sample dataset D constructed in step S2 is as follows:

[0017] D = {(x1,y1),(x2,y2),…,(x n ,y n )}

[0018] In the formula, x m y m Let m and m represent the m-th well logging data sample and its corresponding lithology label, respectively, where m∈{1,2,…,n};

[0019] Based on the CART decision tree algorithm, construct node N and set a threshold. In the formula:

[0020]

[0021] In the formula, V represents the number of lithology categories in the sample dataset D, Gini(D) represents the Gini index, which is the probability that any two well logging data samples in the sample dataset D belong to different lithology categories, and p k p represents the proportion of well logging data samples belonging to the k-th lithology category in the sample dataset D. k ′ represents the proportion of well logging data samples belonging to the k′-th lithology category in the sample dataset D, and k′≠k;

[0022] Step S32: Based on Gini(D), set the node splitting attribute of the CART decision tree algorithm with the smallest Gini index, where the smallest Gini index a * As shown in the following formula:

[0023]

[0024] In the formula, a is a discrete attribute of the sample dataset D;

[0025] Step S33: Set the termination condition as follows:

[0026]

[0027] If the sample dataset D meets the termination condition, then mark node N as a leaf node; otherwise, mark node N as a leaf node. * Generate the next branch node for the split criteria, with each node corresponding to a well logging data sample.

[0028] As a preferred embodiment of the present invention, the specific steps of step S4 are as follows:

[0029] Step S41: Construct the unlabeled dataset U as follows:

[0030]

[0031] In the formula, x m , Let m be the m-th logging data sample and its corresponding pseudo-label, where m∈{1,2,…,n};

[0032] Construct an n×2 matrix S from the unlabeled dataset U. Randomly select an array from the unlabeled dataset U to form an unlabeled data subset U1. Calculate the confidence dΨ of each pseudo-label in the unlabeled data subset U1 according to the query function. d The specific formula is as follows:

[0033]

[0034] Where tr(·) represents the trace operation of a matrix, is the scattering matrix, S B is the between-class scattering matrix, and U / U1 means subtracting the unlabeled data subset U1 from the unlabeled data set U;

[0035] Step S42: Preset the confidence threshold f as follows:

[0036]

[0037] Where Ψ represents a measure for evaluating the confidence of the pseudo-label;

[0038] Construct a candidate sample data set L with well logging data samples where dΨ d < f, and construct a pseudo-label data set P with well logging data samples where dΨ d ≥ f as follows:

[0039]

[0040] Where x m 、 are the m-th well logging data sample and its corresponding pseudo-label respectively, m ∈ {1, 2, …, j}, and j is the total number of well logging data samples in the pseudo-label data set P;

[0041] Step S43: For the candidate sample data set L, based on the preset mapping rule, label the lithology labels for each well logging data sample in the candidate sample data set L one by one, and construct a labeled data set E as follows: <00001​​​​​​​​​​​​​​​​​The BPMN process front-end module is used to generate a workflow for an active learning method for lithology identification based on BPMN according to each logging data sample, and send the workflow in bpmn format to the camunda process engine module.

[0046] The Camunda-based workflow engine module serves as the backend module, parses the workflow object transmitted by the BPMN workflow frontend module, sets and executes the workflow, and transmits the workflow to the lithology identification active learning task module.

[0047] The dataset annotation module is used to annotate lithology labels on well logging data samples and upload them to the storage module, and provide the training set annotation data file path to the dynamic form module.

[0048] The dynamic form module includes a model parameter form module and a data form module. The model parameter form module contains a decision tree architecture, and the editable content of the decision tree architecture includes the selection of tree type, number of trees, number of nodes, and number of rows. The data form module is used for user-defined forms, and the editable content includes the number of lithology categories, input size, and training set annotation data file path.

[0049] The lithology identification active learning module receives the workflow from the Camunda workflow engine module and uses a BPMN-based lithology identification active learning method to perform lithology identification active learning work.

[0050] The storage module includes a data storage module and a configuration file storage module. The data storage module is used to store well logging data samples, and the configuration file storage module is used to store user-defined forms from the dynamic form module.

[0051] The model deployment module is used to deploy the active learning model for lithology identification to the edge device according to the workflow of the received Camunda process engine module.

[0052] As a preferred technical solution of the present invention: the workflow object in the BPMN process front-end module includes start event, end event, service task, user task, script task, data storage module, mutual exclusion gateway, and compatible gateway.

[0053] As a preferred technical solution of the present invention: the active learning module for lithology identification includes a data acquisition module, a data filtering module, a classification model training module, a query function classification module, a candidate set data module, a labeling module, and a model update module;

[0054] The data acquisition module is used to read well logging data samples from the storage module;

[0055] The data filtering module constructs a feature dataset x consisting of well logging data samples based on the well logging data samples from the data reading module, and constructs a lithology tag set y consisting of each lithology tag.

[0056] The classification model training module reads a preset number of well logging data samples from the data filtering module, and constructs a sample dataset D based on each lithology label in the lithology label set y, and builds and trains an active learning initial model for lithology identification.

[0057] The query function classification module reads a preset number of well logging data samples from the data filtering module, uses the lithology identification active learning initial model to assign pseudo-labels to each well logging data sample, and constructs an unlabeled dataset U; based on the query function and a preset confidence threshold, well logging data samples with pseudo-label confidence lower than the preset confidence threshold are transmitted to the candidate set data module; well logging data samples with pseudo-label confidence not lower than the preset confidence threshold are transmitted to the model update module.

[0058] The candidate set data module receives well logging data samples whose pseudo-label credibility is lower than a preset credibility threshold transmitted by the query function classification module, constructs a candidate sample dataset L, and transmits it to the labeling module.

[0059] After obtaining the candidate sample dataset L, the annotation module annotates each well logging data sample in the candidate sample dataset L with a lithological label based on a preset mapping rule, constructs an annotation dataset E, and transmits it to the model update module.

[0060] The model update module receives the datasets from the annotation module, query function classification module, and classification model training module to form a new dataset, and then updates the initial active learning model for lithology identification to form an active learning model for lithology identification.

[0061] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0062] This invention designs a BPMN-based active learning method and system for lithology identification. By adopting a BPMN-based active learning method for lithology identification, it adds a confidence calculation, candidate set extraction, and manual annotation module to the original machine learning steps. Through the active learning of the machine, it improves parameter learning, model training and complexity, enhances model performance, and increases work efficiency. Attached Figure Description

[0063] Figure 1 This is a coded flowchart of the active learning method for lithology identification provided in an embodiment of the present invention;

[0064] Figure 2 This is a flowchart of the BPMN active learning method for lithology identification provided in an embodiment of the present invention;

[0065] Figure 3 This is a structural diagram of the active learning method for lithology identification based on BPMN provided in an embodiment of the present invention;

[0066] Figure 4 This is a flowchart of an active learning method for lithology identification based on BPMN provided in an embodiment of the present invention. Detailed Implementation

[0067] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0068] This invention provides a lithology identification active learning method based on BPMN. For each target well log, the following steps S1-S5 are performed to construct an active learning model for lithology identification, and then the active learning model is applied to complete the identification of the lithology of each target well log:

[0069] Step S1: Based on the preset logging method, collect logging data from each target well, normalize the logging data of each target well to obtain each logging data sample, and construct a feature dataset x consisting of each logging data sample; combine the preset lithology labels corresponding to the preset lithology categories to construct a lithology label set y consisting of each lithology label.

[0070] The preset logging methods in step S1 include one or more of gamma-ray logging, sonic logging, density logging, compensated neutron logging, deep lateral logging, shallow lateral logging, and caliper logging; the lithology labels include acid rocks, coal, sandy conglomerate, coarse sandstone, medium sandstone, fine sandstone, siltstone, and mudstone.

[0071] Step S2: Randomly select n well logging data samples from the feature dataset x, and label each well logging data sample one by one according to each lithology label set y, and construct a sample dataset D from the labeled well logging data samples;

[0072] Step S3: Refer to Figure 1 , Figure 2 Based on the CART decision tree algorithm, a decision tree is divided according to lithology classification to construct an initial active learning model for lithology identification. The sample dataset D constructed in step S2 is used as input, and the lithology label corresponding to each well logging data sample is used as output to train the initial active learning model for lithology identification a preset number of times.

[0073] The specific steps of step S3 are as follows:

[0074] Step S31: The sample dataset D constructed in step S2 is as follows:

[0075] D = {(x1,y1),(x2,y2),…,(x n ,y n )}

[0076] In the formula, x m y m Let m and m represent the m-th well logging data sample and its corresponding lithology label, respectively, where m∈{1,2,…,n};

[0077] Initialize the lithology identification active learning initial model. The initial lithology identification active learning initial model is empty. Set the initialization coefficients, including the number of decision trees T and the number of well logging data samples selected n (n>2). The learning step is initialized to 0. The maximum learning step is a positive integer greater than 1.

[0078] Based on the CART decision tree algorithm, construct node N and set a threshold. In the formula:

[0079]

[0080] In the formula, V represents the number of lithology categories in the sample dataset D, Gini(D) represents the Gini index, which is the probability that any two well logging data samples in the sample dataset D belong to different lithology categories, and p k p represents the proportion of well logging data samples belonging to the k-th lithology category in the sample dataset D. k ′ represents the proportion of well logging data samples belonging to the k′-th lithology category in the sample dataset D, and k′≠k;

[0081] Step S32: Based on Gini(D), set the node splitting attribute of the CART decision tree algorithm with the smallest Gini index, where the smallest Gini index a * As shown in the following formula:

[0082]

[0083] In the formula, a is a discrete attribute of the sample dataset D;

[0084] Step S33: Set the termination condition as follows:

[0085]

[0086] If the sample dataset D meets the termination condition, then mark node N as a leaf node; otherwise, mark node N as a leaf node. * Generate the next branch node for the split criteria, with each node corresponding to a well logging data sample.

[0087] Step S4: Randomly select n logging data samples from the feature dataset x, input each logging data sample into the initial lithology identification active learning model obtained in Step S3, use the lithology labels corresponding to each logging data sample output by the lithology identification active learning model as pseudo-labels, assign each pseudo-label to each logging data sample one by one, and construct an unlabeled dataset U;

[0088] For the unlabeled dataset U, based on the query function and a preset confidence threshold, construct a candidate sample dataset L with the logging data samples whose pseudo-label confidence is lower than the preset confidence threshold, and based on the preset mapping rule, annotate the lithology labels of each logging data sample in the candidate sample dataset L one by one, and construct an annotated dataset E;

[0089] Construct a pseudo-label dataset P with the logging data samples whose pseudo-label confidence is not lower than the preset confidence threshold;

[0090] Step S41: Construct the unlabeled dataset U as follows:

[0091]

[0092] where x m 、 are the m-th logging data sample and its corresponding pseudo-label respectively, m ∈ {1, 2,..., n};

[0093] Form the unlabeled dataset U into an n×2 matrix S, randomly extract an array from the unlabeled dataset U to form an unlabeled data subset U1, and calculate the confidence dΨ of each pseudo-label in the unlabeled data subset U1 according to the query function d , specifically as follows:

[0094]

[0095] where tr(·) represents the trace operation of the matrix, is the scatter matrix, S B is the between-class scatter matrix, and U / U1 means subtracting the unlabeled data subset U1 from the unlabeled dataset U;

[0096] Step S42: Preset the confidence threshold f as follows: <00​​​​​​​​​​​​

[0100]

[0101] In the formula, x m , Let m be the m-th logging data sample and its corresponding pseudo-label, where m∈{1,2,…,j} and j is the total number of logging data samples in the pseudo-label dataset P;

[0102] Step S43: For the candidate sample dataset L, based on the preset mapping rules, label each well logging data sample in the candidate sample dataset L with lithological labels, and construct the labeled dataset E as follows:

[0103] E = {(x1,y1),(x2,y2),…,(x...} i ,y i )}

[0104] In the formula, x m y m Let m and i represent the m-th logging data sample and its corresponding lithology label, respectively, m∈{1,2,…,i}, where i is the total number of logging data samples in the labeled dataset E.

[0105] Step S5: Input the sample dataset D obtained in step S2, the labeled dataset E obtained in step S4, and the pseudo-label dataset P into the initial active learning model for lithology identification, update it to obtain the active learning model for lithology identification, and apply the active learning model for lithology identification to complete the identification of lithology for each target well.

[0106] This invention provides an active learning system for lithology identification based on BPMN, referring to... Figure 2 , Figure 3 The system includes a BPMN workflow front-end module, a dataset annotation module, a lithology identification active learning task module, a dynamic form module, a Camunda-based workflow engine module, a model deployment module, and a storage module, to implement the aforementioned BPMN-based lithology identification active learning method.

[0107] The BPMN process front-end module is used to generate a workflow for an active learning method for lithology identification based on BPMN according to each logging data sample, and send the workflow in bpmn format to the camunda process engine module.

[0108] The Camunda-based workflow engine module serves as the backend module, parses the workflow object transmitted by the BPMN workflow frontend module, sets and executes the workflow, and transmits the workflow to the lithology identification active learning task module.

[0109] The dataset annotation module is used to annotate lithology labels on well logging data samples and upload them to the storage module, and provide the training set annotation data file path to the dynamic form module.

[0110] The dynamic form module is used for user-defined forms and saves form data in YAML format in the storage module. It includes a model parameter form module and a data form module. The model parameter form module contains a decision tree architecture. The editable content of the decision tree architecture includes the selection of tree type, number of trees, number of nodes, and number of rows. The data form module is used for user-defined forms and the editable content includes the number of lithology categories, input size, and training set annotation data file path.

[0111] The data form module contains the following information, as shown in Table 1: the editable data forms include the number of lithology categories, input size, training set annotation data file path, and their formats.

[0112] Table 1

[0113]

[0114] The data form content of the model parameter form module is shown in Table 2, which includes decision tree architecture and RF tree architecture. The editable content of the decision tree architecture includes tree type selection, number of trees, number of nodes, and number of rows.

[0115] Table 2

[0116]

[0117] The lithology identification active learning module receives a workflow from the Camunda workflow engine module and employs a BPMN-based lithology identification active learning method to perform lithology identification active learning. The lithology identification active learning task module executes learning tasks according to user-defined algorithm functions. When the task flow reaches a specific learning task, the machine active learning workflow is executed. When a background script task triggers the model deployment module, the model is deployed to the machine device, and the active learning service is started on the machine device via a script.

[0118] Reference Figure 4 This is a flowchart of the lithology identification active learning system based on BPMN of the present invention. It includes the following steps:

[0119] Step 1: The user designs the active learning process for lithology identification, including service task design, user task design, script task design, and gateway design. The data to be learned and the model architecture are selected through the design of the database.

[0120] Step 2: The user generates a BPMN process through the BPMN process front-end module and sends it to the Camunda-based process engine module to generate a workflow. The workflow includes the following tasks:

[0121] 1. Active learning workflow task: Obtain tasks in the workflow. When the corresponding module executes the corresponding task, it triggers the corresponding active learning workflow task. The learning task is executed according to the algorithm function proposed by the user. When the task flow reaches the learning task, the corresponding machine active learning workflow task is executed.

[0122] 2. When the workflow reaches the user service stage, it waits for user execution. User tasks include uploading the labeled and organized data to the storage module, and manually labeling the data after supervised learning. After labeling, the user selects "complete task," and the workflow begins the next operation.

[0123] 3. When the background workflow reaches the script service, the script is executed. The script service triggers the model deployment module to deploy the model to the machine device, and the active learning task is started on the machine device via the script.

[0124] The storage module uses MinIO object storage for storage, including a data storage module and a configuration file storage module. The data storage module is used to save well logging data samples, and the saving format is various well logging data. The configuration file storage module is used to save user-defined forms from the dynamic form module and user-defined lithology identification learning models, and the saving format is a pth file.

[0125] The model deployment module is used to deploy the active learning model for lithology identification to edge devices according to the workflow received from the Camunda workflow engine module. When the model deployment module receives the workflow from the Camunda workflow engine module, it combines the sub-processes in the BPMN with related task activities. When the workflow executes the corresponding script task, the script task triggers the model deployment module. The server uses the TCP / IP protocol as the connection protocol between the server and the device. Different edge devices on the client side send connection requests to the server based on the server's IP address and port. After the server receives the request and successfully connects, it transmits the active learning model for lithology identification to edge devices such as the Jetson Nano, thus achieving model deployment to edge devices such as the Jetson Nano.

[0126] The workflow objects in the BPMN workflow front-end module include start events, end events, service tasks, user tasks, script tasks, data storage modules, mutual exclusion gateways, and compatible gateways.

[0127] The active learning module for lithology identification includes a data acquisition module, a data filtering module, a classification model training module, a query function classification module, a candidate set data module, a labeling module, and a model update module.

[0128] The data acquisition module is used to read well logging data samples from the storage module;

[0129] The data filtering module constructs a feature dataset x consisting of well logging data samples based on the well logging data samples from the data reading module, and constructs a lithology tag set y consisting of each lithology tag.

[0130] The classification model training module reads a preset number of well logging data samples from the data filtering module, and constructs a sample dataset D based on each lithology label in the lithology label set y, and builds and trains an active learning initial model for lithology identification.

[0131] The query function classification module reads a preset number of well logging data samples from the data filtering module, uses the lithology identification active learning initial model to assign pseudo-labels to each well logging data sample, and constructs an unlabeled dataset U; based on the query function and a preset confidence threshold, well logging data samples with pseudo-label confidence lower than the preset confidence threshold are transmitted to the candidate set data module; well logging data samples with pseudo-label confidence not lower than the preset confidence threshold are transmitted to the model update module.

[0132] The candidate set data module receives well logging data samples whose pseudo-label credibility is lower than a preset credibility threshold transmitted by the query function classification module, constructs a candidate sample dataset L, and transmits it to the labeling module.

[0133] After obtaining the candidate sample dataset L, the annotation module annotates each well logging data sample in the candidate sample dataset L with a lithological label based on a preset mapping rule, constructs an annotation dataset E, and transmits it to the model update module.

[0134] The model update module receives the datasets from the annotation module, query function classification module, and classification model training module to form a new dataset, and then updates the initial active learning model for lithology identification to form an active learning model for lithology identification.

[0135] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A lithology identification active learning method based on BPMN, characterized in that, For each target well log, perform the following steps S1-S5 to construct an active learning model for lithology identification, and apply the active learning model to complete the lithology identification for each target well log: Step S1: Based on the preset logging method, collect logging data from each target well, normalize the logging data of each target well to obtain each logging data sample, and construct a feature dataset x consisting of each logging data sample; combine the preset lithology labels corresponding to the preset lithology categories to construct a lithology label set y consisting of each lithology label. Step S2: Randomly select n well logging data samples from the feature dataset x, and label each well logging data sample one by one according to each lithology label set y, and construct a sample dataset D from the labeled well logging data samples; Step S3: Based on the CART decision tree algorithm, divide the decision tree according to lithology classification and construct an initial active learning model for lithology identification; Using the sample dataset D constructed in step S2 as input and the lithology label corresponding to each well logging data sample as output, the lithology identification active learning initial model is trained a preset number of times. Step S4: Randomly select n well logging data samples from the feature dataset x, input each well logging data sample into the lithology identification active learning initial model obtained in step S3, use the lithology label corresponding to each well logging data sample output by the lithology identification active learning initial model as pseudo label, assign each pseudo label to each well logging data sample one by one, and construct an unlabeled dataset U. For the unlabeled dataset U, based on the query function and the preset confidence threshold, a candidate sample dataset L is constructed using well logging data samples with pseudo-label confidence lower than the preset confidence threshold. Based on the preset mapping rules, each well logging data sample in the candidate sample dataset L is labeled with a lithological label, and a labeled dataset E is constructed. A pseudo-label dataset P is constructed using well logging data samples with pseudo-label credibility not lower than a preset credibility threshold; Step S5: Input the sample dataset D obtained in step S2, the labeled dataset E obtained in step S4, and the pseudo-label dataset P into the initial active learning model for lithology identification, update it to obtain the active learning model for lithology identification, and apply the active learning model for lithology identification to complete the identification of lithology for each target well.

2. The active learning method for lithology identification based on BPMN according to claim 1, characterized in that, The preset logging methods in step S1 include one or more of gamma-ray logging, sonic logging, density logging, compensated neutron logging, deep lateral logging, shallow lateral logging, and caliper logging; the lithology labels include acid rocks, coal, sandy conglomerate, coarse sandstone, medium sandstone, fine sandstone, siltstone, and mudstone.

3. The active learning method for lithology identification based on BPMN according to claim 1, characterized in that, The specific steps of step S3 are as follows: Step S31: The sample dataset D constructed in step S2 is as follows: D={(x1,y1),(x2,y2),…,(x n ,y n )} In the formula, x m y m Let m and m represent the m-th well logging data sample and its corresponding lithology label, respectively, where m∈{1,2,…,n}; Based on the CART decision tree algorithm, construct node N and set a threshold. In the formula: In the formula, V represents the number of lithology categories in the sample dataset D, Gini(D) represents the Gini index, which is the probability that any two well logging data samples in the sample dataset D belong to different lithology categories, and p k p represents the proportion of well logging data samples belonging to the k-th lithology category in the sample dataset D. k ′ represents the proportion of well logging data samples belonging to the k′-th lithology category in the sample dataset D, and k′≠k; Step S32: Based on Gini(D), set the node splitting attribute of the CART decision tree algorithm with the smallest Gini index, where the smallest Gini index a * As shown in the following formula: In the formula, a is a discrete attribute of the sample dataset D; Step S33: Set the termination condition as follows: If the sample dataset D meets the termination condition, then mark node N as a leaf node; otherwise, mark node N as a leaf node. * Generate the next branch node for the split criteria, with each node corresponding to a well logging data sample.

4. The active learning method for lithology identification based on BPMN according to claim 3, characterized in that, The specific steps of step S4 are as follows: Step S41: Construct the unlabeled dataset U as follows: In the formula, x m , Let m be the m-th logging data sample and its corresponding pseudo-label, where m∈{1,2,…,n}; Construct an n×2 matrix S from the unlabeled dataset U. Randomly select an array from the unlabeled dataset U to form an unlabeled data subset U1. Calculate the confidence dΨ of each pseudo-label in the unlabeled data subset U1 according to the query function. d The specific formula is as follows: In the formula, tr(·) represents the matrix tracking operation. Let S be the scattering matrix. B U / U1 is the inter-class scattering matrix, where U / U1 represents the unlabeled dataset U minus the unlabeled data subset U1. Step S42: The preset confidence threshold f is as follows: In the formula, Ψ represents the measure of the credibility of the pseudo-label; With dΨ d <Construct a candidate sample dataset L from the logging data samples of f, with dΨ d ≥Construct a pseudo-label dataset P from the logging data samples of f as follows: In the formula, x m , Let m be the m-th logging data sample and its corresponding pseudo-label, where m∈{1,2,…,j} and j is the total number of logging data samples in the pseudo-label dataset P; Step S43: For the candidate sample dataset L, based on the preset mapping rules, label each well logging data sample in the candidate sample dataset L with lithological labels, and construct the labeled dataset E as follows: E={(x1,y1),(x2,y2),…,(x i ,y i )} In the formula, x m y m Let m and i represent the m-th logging data sample and its corresponding lithology label, respectively, m∈{1,2,…,i}, where i is the total number of logging data samples in the labeled dataset E.

5. A lithology identification active learning system based on BPMN, characterized in that, It includes a BPMN workflow front-end module, a dataset annotation module, a lithology identification active learning task module, a dynamic form module, a Camunda-based workflow engine module, a model deployment module, and a storage module, to implement a BPMN-based lithology identification active learning method as described in any one of claims 1-4. The BPMN process front-end module is used to generate a workflow for an active learning method for lithology identification based on BPMN according to each logging data sample, and send the workflow in bpmn format to the camunda process engine module. The Camunda-based workflow engine module serves as the backend module, parses the workflow object transmitted by the BPMN workflow frontend module, sets and executes the workflow, and transmits the workflow to the lithology identification active learning task module. The dataset annotation module is used to annotate lithology labels on well logging data samples and upload them to the storage module, and provide the training set annotation data file path to the dynamic form module. The dynamic form module includes a model parameter form module and a data form module. The model parameter form module contains a decision tree architecture, and the editable content of the decision tree architecture includes the selection of tree type, number of trees, number of nodes, and number of rows. The data form module is used for user-defined forms, and the editable content includes the number of lithology categories, input size, and training set annotation data file path. The lithology identification active learning module receives the workflow from the Camunda workflow engine module and uses a BPMN-based lithology identification active learning method to perform lithology identification active learning work. The storage module includes a data storage module and a configuration file storage module. The data storage module is used to store well logging data samples, and the configuration file storage module is used to store user-defined forms from the dynamic form module. The model deployment module is used to deploy the active learning model for lithology identification to the edge device according to the workflow of the received Camunda process engine module.

6. The lithology identification active learning system based on BPMN according to claim 5, characterized in that, The workflow objects in the BPMN workflow front-end module include start events, end events, service tasks, user tasks, script tasks, data storage modules, mutual exclusion gateways, and compatible gateways.

7. The lithology identification active learning system based on BPMN according to claim 5, characterized in that, The active learning module for lithology identification includes a data acquisition module, a data filtering module, a classification model training module, a query function classification module, a candidate set data module, a labeling module, and a model update module. The data acquisition module is used to read well logging data samples from the storage module; The data filtering module constructs a feature dataset x consisting of well logging data samples based on the well logging data samples from the data reading module, and constructs a lithology tag set y consisting of each lithology tag. The classification model training module reads a preset number of well logging data samples from the data filtering module, and constructs a sample dataset D based on each lithology label in the lithology label set y, and builds and trains an active learning initial model for lithology identification. The query function classification module reads a preset number of well logging data samples from the data filtering module, uses the lithology identification active learning initial model to assign pseudo-labels to each well logging data sample, and constructs an unlabeled dataset U; based on the query function and a preset confidence threshold, well logging data samples with pseudo-label confidence lower than the preset confidence threshold are transmitted to the candidate set data module; well logging data samples with pseudo-label confidence not lower than the preset confidence threshold are transmitted to the model update module. The candidate set data module receives well logging data samples whose pseudo-label credibility is lower than a preset credibility threshold transmitted by the query function classification module, constructs a candidate sample dataset L, and transmits it to the labeling module. After obtaining the candidate sample dataset L, the annotation module annotates each well logging data sample in the candidate sample dataset L with a lithological label based on a preset mapping rule, constructs an annotation dataset E, and transmits it to the model update module. The model update module receives the datasets from the annotation module, query function classification module, and classification model training module to form a new dataset, and then updates the initial active learning model for lithology identification to form an active learning model for lithology identification.

Images